Pipeline and Backpressure

Ranya uses a staged pipeline that runs per call. The pipeline accepts frames, routes them through processors, and emits frames to the transport.

Default Processor Order

The default engine builds this sequence in pkg/ranya/engine.go:

1. Pre‑processors (acoustic, network simulation).
2. STT processor.
3. Turn processor.
4. Before‑context processors.
5. Context processor.
6. Router processor.
7. Before‑LLM processors.
8. LLM processor.
9. Tool dispatcher.
10. Before‑TTS processors.
11. TTS processor.
12. Post‑processors (serializers).

Custom apps can add or remove stages through EngineOptions.

Backpressure Modes

• pipeline.backpressure=drop drops frames when a channel is full.
• pipeline.backpressure=wait blocks until space is available.

When to use each:

• Drop for live calls where latency matters more than completeness.
• Wait for offline or test runs where you want every frame.

Tuning Recipes

Goal	Settings
Lowest latency	backpressure=drop, smaller buffers.
Highest reliability	backpressure=wait, larger buffers.
Debug deterministically	pipeline.async=false.
Handle heavy load	increase highcapacity, lowcapacity, stagebuffer.

Diagram

flowchart LR In[Input Frames] --> PQ[Priority Queue] PQ --> S0[Stage 0] S0 --> S1[Stage 1] S1 --> S2[Stage N] S2 --> Out[Output Frames] PQ --> H[High Control] PQ --> L[Low Audio and Text]